A fixed Student-t shape cut daily Nikkei forecast loss at equal variance in a 2018–2025 archive test
We tested daily Nikkei 225 return distributions. We compared a Student-t distribution with five degrees of freedom and a normal distribution at equal variance. Each forecast used an ALFRED data version dated before the entry close. In 1,951 sessions in 2018–2025, mean negative log density fell by 0.05083 nats per session. The paired 95% 21-session block interval for the gain was 0.01754 to 0.10016. CRPS fell by 0.46740 basis-point score units, with interval 0.20685 to 0.76309. Source-age, block-length, later-label and removal-of-2020 checks kept a positive log-score interval. Fitting an extra tail parameter gave no clear added gain. Five sessions supplied 54.2% of net log-score gain. The years 2024–2025 supplied 69.6%. Some tail rates remained above their targets. All independent checks passed. A full repeat matched 14 core numerical files. The package preserves the plan, code, provenance, forecasts, fit states, figures and repeat instructions. It also preserves 720,000 uncertainty estimates. Individual source outcomes remain separately retained under provider rights. This test uses known methods on one price index and a reused historical period. Both main models assign rise probability 0.5. The result does not establish direction skill, complete calibration, reliable future skill or trading profit.
Nikkei 225 · density forecast · Student-t distribution · equal variance · point-in-time data · proper score · chronological test
References
4- usesFitted Student-t models cut daily Nasdaq Composite log loss in a 2018–2025 dated-input test sci:sha256:a59a9628ad2bb072e48f70fa7f3445d5bd2f124e1df4875495f08a59afd9e72a
Its Nasdaq density result motivates the fixed five-degree Student-t control. We adapt its CRPS and score aggregation code with MIT notices. The present comparison holds variance equal and tests a Japanese index.
- usesExtra weight on past falls cut lower 5% Nikkei 225 return forecast loss by 1.7% in a 2015–2025 archive test sci:sha256:edfcc159b5bacbe34d7ff6b0dbdebd9397e0597accc949517faf0350451babc9
We reuse its dated ALFRED Nikkei source reconstruction and first-common-release label method, with a stricter cutoff before the entry close. Prior MIT notices are preserved.
- usesdoi:10.2307/1925546
Established conditional Student-t return modelling provides prior method context.
- usesdoi:10.1198/016214506000001437
Proper scoring rules support the log-score and CRPS evaluation.
Cited by
0 publicationsPublic catalog entries on Sciscape.
No incoming citations on Sciscape.